Hyperspectral image (HSI) analysis is a growing area in the community of remote sensing, particularly with images exhibiting high spatial and spectral resolutions. Multiple kernel learning (MKL) has been proposed and found to classify HSIs efficiently owing to its capability for handling diverse feature fusion. However, constructing base kernels, selecting key kernels, and adjusting their contributions to the final kernel remain major challenges for MKL. We propose a scheme to generate effective base kernels and optimize their weights, which represent their contribution to the final kernel. In addition, both spatial and spectral information are utilized to improve the classification accuracy. In the proposed scheme, the spatial features of HSIs are introduced through multiscale feature representations that preserve the relationship between the classification process and the pixel context. MKL and self-organizing maps (SOMs) are integrated and used for the unsupervised classification of HSIs. The weights of both the base kernels and neural networks are simultaneously optimized in an unsupervised manner. The results indicate that the proposed MKL-SOM scheme outperforms state-of-the-art algorithms, particularly when applied to large HSIs. Moreover, its ability to fuse multiscale features, especially in large HSIs, is useful for various analysis tasks.